EMDKG: Improving Accuracy-Diversity Trade-Off in Recommendation with EM-based Model and Knowledge Graph Embedding

To maintain attractiveness and reduce redundancy of recommendation, the concept of diversity has been brought up in recommender systems (RS). Thus, advanced RS aim at achieving both better accuracy and diversity facing a trade-off issue between the two aspects. Recently, knowledge graphs embedding methods have been widely used in RS for achieving better accuracy provided with auxiliary information along with historical user-item interactions. However, little work has been done to investigate what effects of diversity it brings along with higher accuracy results and how to achieve the best accuracy-diversity trade-off under such circumstances. In this paper, we propose an EM-model capable of incorporating a generalized concept of diversity for a diversity-encoded knowledge graph embedding based recommendation. Our EM-model alternates between a general item diversity learning and knowledge graph embedding learning for user and item representation, which helps to achieve better results in terms of both accuracy and diversity compared to the state-of-art baselines on datasets MovieLens and Anime. Moreover, extensive experiments prove our model outperforms the baseline with existing diversification methods (MMR and DPP) achieving a better accuracy-diversity trade-off.